Detection and Tracking Human · PDF fileModeling techniques Applications Summary ......
Transcript of Detection and Tracking Human · PDF fileModeling techniques Applications Summary ......
Human Motion Detection and TrackingA Technical Presentation
Members● William Juszczyk● Michael Lazar● Ryan Lattrel● Mark Birdsall● Camden Smith
Design Team 2Moving Human Electromagnetic
Scattering SimulatorFacilitator● Dr. Wen Li
Sponsor● Air Force Research
Laboratory
● Introduction● History● Methods of detection● Modeling techniques● Applications● Summary● Questions
Presentation Outline
● Human motion tracking is a subset motion tracking
● Accurately identifying human activity has become important in a large number of fields
● Developments in camera technologies and data processing make complex system possible
Introduction
History ● Infrared Search and Track (IRST) - First IR Detection● Motion Controllers
a. SEGA Activator - first Motion Controller (1993)b. Nintendo Wii-Remote - using accelerometers
(2006)● Depth Sensing Technology,
Xbox Kinect first systemto use DST.
Picture - SEGA Activator. Sits on the groundwhile the player stands in middle.
http://mytechroom.co.uk/kinectivator/
Methods of Detection
Methods of Detection● There are many ways to detect and track
humans○ Sound○ Vibration○ Ultrasonic sonar○ Infrared
● All have unique advantages and disadvantages
Sound● Monitor acoustic
sound signatures● Footsteps have
broadband frequency○ Generated by foot
striking the ground● Works best
indoors(http://www.tpub.com/neets/book10/NTX1-14.GIF)
Vibration● Detection with an
accelerometer● Difficult to analyze
○ Varies with site● Can work in non-
ideal conditions○ Outdoors○ Wind or noise
(https://en.wikipedia.org/wiki/File:Kinemetrics_seismograph.jpg)
Ultrasonic Sonar● Uses sound waves● Can detect moving
objects● Wide field of vision
○ Infrared is more limited
○ Sound can travel further
(http://www.cypress.com/?rID=82)
Infrared● Detect emitted
energy○ Unlike ultrasonic
sonar, there is no transmitter
● Sees changes of temperature around a point
● Can be calibrated to track movement
(https://upload.wikimedia.org/wikipedia/commons/2/25/ParowozIR.jpg)
Microsoft Modeling
Algorithm Approaches● Calibration pose
○ Old theory○ Problems?
● Microsoft approach○ Robust solution○ Scan 3D image○ This research led to Kinect SDK
http://leonardodavinci.stanford.edu/submissions/clabaugh/history/leonardo.html
Joint Labeling● Single depth
image● Body part labeling● 3D joint proposals● Body parts are
then treated per pixel http://research.microsoft.
com/pubs/145347/BodyPartRecognition.pdf
How Decisions are Made● Generate "realistic" depth images● Use a decision forest
○ A decision forest is a group of decision trees● Speed issues
○ Each pixel run in parallel on the decision forest○ On Xbox 360 GPU: 5ms per frame
http://research.microsoft.com/pubs/145347/BodyPartRecognition.pdf
Alternative Modeling
2D Model● biped model● 6 rectangles in the model
○ head○ torso○ left/right thigh○ left/right calf
● Detection and tracking algorithm in a 3 step loop○ Detecting human candidate○ Validating model of a human○ Tracking of the model in subsequent frames
Detecting Human Candidate● Input is output of a pixel-based
motion detector● A rectangular mask is compared
to the different parts of the image○ If there is motion detected that can be
fit within the mask, then this is a region of interest (ROI)
○ The mask is scaled based on vertical location of ROI
● Assumes people are roughly the same height
http://www.sciencedirect.com/science/article/pii/S0031320308000071
Validating the Model● Biped model is initialized at each ROI
○ Determines position and orientation of model
○ Pose of model that fits ROI best is calculated using a mean-shift algorithm
○ Location of torso found first, then legs, then head
● Model is refined○ Vertical position of each model part needs
to be shifted to better fit the image○ Non-humans will be thrown out.
http://cmp.felk.cvut.cz/demos/Tracking/TrackHumansKorc/
Tracking of Model● Pose of model from previous frame is kept● Algorithm tries to anticipate the human pose
in the current frame based on the previous frame.
● Runs the same algorithm from the Validation phase to fit model to image.
● There are limitations due to occlusion.
Applications
● Automotive○ Motivated by Safety○ MobilEye PCW○ Warning Systems○ Active vs. Passive
● Surveillance○ More Efficient○ Improved reliability○ Used in airports, subway stations, office buildings,
private use
Applications
http://www.mobileye.com/technology/applications/pedestrian-detection/pedestrian-collision-warning/
● Content based video categorization● Sports video analysis● Motion capture for film and television● Gesture recognition for smart devices
More Applications
Summary● History
○ Infrared ○ motion controllers
● Methods of Detection○ Sound, vibration, sonar, and infrared
● Modeling Humans○ Microsoft's approach○ Two dimensional model
● Applications○ Auto○ Security
Questions?
ReferencesMotion Detection History: 1) http://illumin.usc.edu/165/motion-sensors/2) http://en.wikipedia.org/wiki/Infra-red_search_and_trackMethods:1) http://www.doityourself.com/stry/ultrasonic-vs-infrared-wireless-motion-detectors2) http://www.cypress.com/?rID=823) https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5935481&tag=14) http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA4787955) http://research.microsoft.com/pubs/145347/BodyPartRecognition.pdf